DeepBatch: A hybrid deep learning model for interpretable diagnosis of breast cancer in whole-slide images

Abstract

The gold standard for breast cancer diagnosis, treatment, and management is the histological analysis of a suspected section. Histopathology consists in analyzing the characteristics of the lesions using tissue sections stained with hematoxylin and eosin. However, pathologists are currently subjected to high workloads, mainly due to the fundamental role of histological analysis in the patient’s treatment. In this context, methods able to reduce histological analysis time, provide a second opinion, or even point out suspicious locations as a screening tool become increasingly important for pathologists. This article proposes a model based on Convolutional Neural Networks (CNN) to provide a refined and multiclass segmentation of Whole Slide Imaging (WSI) for breast cancer. The methodology is divided into four modules: pre-processing, ROI detection, ROI sampling, and region segmentation. These modules are organized to decode the information learned using CNNs in interpretable predictions for pathologists. The preprocessing module is responsible for removing background and noise from WSI. At ROI detection, we use the U-Net convolutional architecture to identify suspicious regions in low magnification WSI. The sampling module maps the identified suspected areas from low magnifications to 40× magnifications. region segmentation module segments high-magnification areas using a ResNet50/U-Net. To validate the methodology, we use data sets from different sources that can be used together or separately in each module, depending on its purpose. We used 205 breast cancer WSI for training, validation, and testing. For the detection of suspicious regions by ROI detection, we obtained an IoU of 93.43%, accuracy of 91.27%, sensitivity of 90.77%, specificity of 94.03%, F1 score of 84.17%, and an AUC of 0.93. For the refined segmentation of WSI by the region segmentation module, we obtained an IoU of 88.23%, accuracy of 96.10%, sensitivity of 71.83%, specificity of 96.19%, F1 score of 82.94%, and an AUC of 0.88. In short, the model provides refined segmentation of breast cancer WSIs using a cascade of CNNs. This segmentation can assist pathologists in interpreting the diagnosis by accurately presenting the regions considered during the inference of WSI. Our results indicate the possibility of using the model as a second screening system.